Weighted Strategy Optimization Approach for Discrete Sequence Generation

Generative Adversarial Networks (GANs) perform well in continuous data generation, but face challenges in discrete sequence generation due to the non-differentiability of backpropagation. SeqGAN addresses this issue by utilizing the policy gradient method; however, it still suffers from limitations such as low efficiency, inadequate reward utilization, and a tendency to converge to local optima. This paper proposes a discrete sequence generation model, W-SeqGAN, which is based on Weighted Policy Optimization (WPO). This model incorporates WPO into the training process of the generator. Its dynamic weighting mechanism ena bles the generator to focus on samples with strong discriminator feedback or those that are difficult to generate, thereby enhancing the generation quality. Experiments were conducted using synthe tic data, and the Negative Log Likelihood (NLL) was adopted as t he evaluation metric. The results indicate that W-SeqGAN outper forms traditional SeqGAN and other baseline models in terms of generation quality and training efficiency.